Beyond Compression: Training Latent Representations for Stable Long-Horizon Rollout in Neural Surrogate Solvers — ThinkLLM